1 #pragma once 2 3 #include <petscdevice.h> 4 #include <petscpkg_version.h> 5 6 /* MANSEC = Sys */ 7 8 #if defined(__NVCC__) || defined(__CUDACC__) 9 #define PETSC_USING_NVCC 1 10 #endif 11 12 #if PetscDefined(HAVE_CUDA) 13 #include <cuda.h> 14 #include <cuda_runtime.h> 15 #include <cublas_v2.h> 16 #define DISABLE_CUSPARSE_DEPRECATED 17 #include <cusparse.h> 18 #include <cusolverDn.h> 19 #include <cusolverSp.h> 20 #include <cufft.h> 21 #include <curand.h> 22 #include <nvml.h> // NVML comes with the NVIDIA GPU driver 23 24 /* cuBLAS does not have cublasGetErrorName(). We create one on our own. */ 25 PETSC_EXTERN const char *PetscCUBLASGetErrorName(cublasStatus_t); /* PETSC_EXTERN since it is exposed by the CHKERRCUBLAS macro */ 26 PETSC_EXTERN const char *PetscCUSolverGetErrorName(cusolverStatus_t); 27 PETSC_EXTERN const char *PetscCUFFTGetErrorName(cufftResult); 28 29 /* REMOVE ME */ 30 #define WaitForCUDA() cudaDeviceSynchronize() 31 32 /* CUDART_VERSION = 1000 x major + 10 x minor version */ 33 34 /* Could not find exactly which CUDART_VERSION introduced cudaGetErrorName. At least it was in CUDA 8.0 (Sep. 2016) */ 35 #if PETSC_PKG_CUDA_VERSION_GE(8, 0, 0) 36 #define PetscCallCUDAVoid(...) \ 37 do { \ 38 const cudaError_t _p_cuda_err__ = __VA_ARGS__; \ 39 PetscCheckAbort(_p_cuda_err__ == cudaSuccess, PETSC_COMM_SELF, PETSC_ERR_GPU, "cuda error %d (%s) : %s", (PetscErrorCode)_p_cuda_err__, cudaGetErrorName(_p_cuda_err__), cudaGetErrorString(_p_cuda_err__)); \ 40 } while (0) 41 42 #define PetscCallCUDA(...) \ 43 do { \ 44 const cudaError_t _p_cuda_err__ = __VA_ARGS__; \ 45 PetscCheck(_p_cuda_err__ == cudaSuccess, PETSC_COMM_SELF, PETSC_ERR_GPU, "cuda error %d (%s) : %s", (PetscErrorCode)_p_cuda_err__, cudaGetErrorName(_p_cuda_err__), cudaGetErrorString(_p_cuda_err__)); \ 46 } while (0) 47 #else /* PETSC_PKG_CUDA_VERSION_GE(8,0,0) */ 48 #define PetscCallCUDA(...) \ 49 do { \ 50 const cudaError_t _p_cuda_err__ = __VA_ARGS__; \ 51 PetscCheck(_p_cuda_err__ == cudaSuccess, PETSC_COMM_SELF, PETSC_ERR_GPU, "cuda error %d", (PetscErrorCode)_p_cuda_err__); \ 52 } while (0) 53 54 #define PetscCallCUDAVoid(...) \ 55 do { \ 56 const cudaError_t _p_cuda_err__ = __VA_ARGS__; \ 57 PetscCheckAbort(_p_cuda_err__ == cudaSuccess, PETSC_COMM_SELF, PETSC_ERR_GPU, "cuda error %d", (PetscErrorCode)_p_cuda_err__); \ 58 } while (0) 59 #endif /* PETSC_PKG_CUDA_VERSION_GE(8,0,0) */ 60 #define CHKERRCUDA(...) PetscCallCUDA(__VA_ARGS__) 61 62 #define PetscCUDACheckLaunch \ 63 do { \ 64 /* Check synchronous errors, i.e. pre-launch */ \ 65 PetscCallCUDA(cudaGetLastError()); \ 66 /* Check asynchronous errors, i.e. kernel failed (ULF) */ \ 67 PetscCallCUDA(cudaDeviceSynchronize()); \ 68 } while (0) 69 70 #define PetscCallCUBLAS(...) \ 71 do { \ 72 const cublasStatus_t _p_cublas_stat__ = __VA_ARGS__; \ 73 if (PetscUnlikely(_p_cublas_stat__ != CUBLAS_STATUS_SUCCESS)) { \ 74 const char *name = PetscCUBLASGetErrorName(_p_cublas_stat__); \ 75 if (((_p_cublas_stat__ == CUBLAS_STATUS_NOT_INITIALIZED) || (_p_cublas_stat__ == CUBLAS_STATUS_ALLOC_FAILED)) && PetscDeviceInitialized(PETSC_DEVICE_CUDA)) { \ 76 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU_RESOURCE, \ 77 "cuBLAS error %d (%s). " \ 78 "Reports not initialized or alloc failed; " \ 79 "this indicates the GPU may have run out resources", \ 80 (PetscErrorCode)_p_cublas_stat__, name); \ 81 } else { \ 82 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU, "cuBLAS error %d (%s)", (PetscErrorCode)_p_cublas_stat__, name); \ 83 } \ 84 } \ 85 } while (0) 86 #define CHKERRCUBLAS(...) PetscCallCUBLAS(__VA_ARGS__) 87 88 #if (CUSPARSE_VER_MAJOR > 10 || CUSPARSE_VER_MAJOR == 10 && CUSPARSE_VER_MINOR >= 2) /* According to cuda/10.1.168 on OLCF Summit */ 89 #define PetscCallCUSPARSE(...) \ 90 do { \ 91 const cusparseStatus_t _p_cusparse_stat__ = __VA_ARGS__; \ 92 if (PetscUnlikely(_p_cusparse_stat__)) { \ 93 const char *name = cusparseGetErrorName(_p_cusparse_stat__); \ 94 const char *descr = cusparseGetErrorString(_p_cusparse_stat__); \ 95 PetscCheck((_p_cusparse_stat__ != CUSPARSE_STATUS_NOT_INITIALIZED) && (_p_cusparse_stat__ != CUSPARSE_STATUS_ALLOC_FAILED), PETSC_COMM_SELF, PETSC_ERR_GPU_RESOURCE, \ 96 "cuSPARSE errorcode %d (%s) : %s.; " \ 97 "this indicates the GPU has run out resources", \ 98 (int)_p_cusparse_stat__, name, descr); \ 99 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU, "cuSPARSE errorcode %d (%s) : %s", (int)_p_cusparse_stat__, name, descr); \ 100 } \ 101 } while (0) 102 #else /* (CUSPARSE_VER_MAJOR > 10 || CUSPARSE_VER_MAJOR == 10 && CUSPARSE_VER_MINOR >= 2) */ 103 #define PetscCallCUSPARSE(...) \ 104 do { \ 105 const cusparseStatus_t _p_cusparse_stat__ = __VA_ARGS__; \ 106 PetscCheck(_p_cusparse_stat__ == CUSPARSE_STATUS_SUCCESS, PETSC_COMM_SELF, PETSC_ERR_GPU, "cuSPARSE errorcode %d", (PetscErrorCode)_p_cusparse_stat__); \ 107 } while (0) 108 #endif /* (CUSPARSE_VER_MAJOR > 10 || CUSPARSE_VER_MAJOR == 10 && CUSPARSE_VER_MINOR >= 2) */ 109 #define CHKERRCUSPARSE(...) PetscCallCUSPARSE(__VA_ARGS__) 110 111 #define PetscCallCUSOLVER(...) \ 112 do { \ 113 const cusolverStatus_t _p_cusolver_stat__ = __VA_ARGS__; \ 114 if (PetscUnlikely(_p_cusolver_stat__ != CUSOLVER_STATUS_SUCCESS)) { \ 115 const char *name = PetscCUSolverGetErrorName(_p_cusolver_stat__); \ 116 if (((_p_cusolver_stat__ == CUSOLVER_STATUS_NOT_INITIALIZED) || (_p_cusolver_stat__ == CUSOLVER_STATUS_ALLOC_FAILED) || (_p_cusolver_stat__ == CUSOLVER_STATUS_INTERNAL_ERROR)) && PetscDeviceInitialized(PETSC_DEVICE_CUDA)) { \ 117 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU_RESOURCE, \ 118 "cuSolver error %d (%s). " \ 119 "This indicates the GPU may have run out resources", \ 120 (PetscErrorCode)_p_cusolver_stat__, name); \ 121 } else { \ 122 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU, "cuSolver error %d (%s)", (PetscErrorCode)_p_cusolver_stat__, name); \ 123 } \ 124 } \ 125 } while (0) 126 #define CHKERRCUSOLVER(...) PetscCallCUSOLVER(__VA_ARGS__) 127 128 #define PetscCallCUFFT(...) \ 129 do { \ 130 const cufftResult_t _p_cufft_stat__ = __VA_ARGS__; \ 131 if (PetscUnlikely(_p_cufft_stat__ != CUFFT_SUCCESS)) { \ 132 const char *name = PetscCUFFTGetErrorName(_p_cufft_stat__); \ 133 if (((_p_cufft_stat__ == CUFFT_SETUP_FAILED) || (_p_cufft_stat__ == CUFFT_ALLOC_FAILED)) && PetscDeviceInitialized(PETSC_DEVICE_CUDA)) { \ 134 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU_RESOURCE, \ 135 "cuFFT error %d (%s). " \ 136 "Reports not initialized or alloc failed; " \ 137 "this indicates the GPU has run out resources", \ 138 (PetscErrorCode)_p_cufft_stat__, name); \ 139 } else { \ 140 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU, "cuFFT error %d (%s)", (PetscErrorCode)_p_cufft_stat__, name); \ 141 } \ 142 } \ 143 } while (0) 144 #define CHKERRCUFFT(...) PetscCallCUFFT(__VA_ARGS__) 145 146 #define PetscCallCURAND(...) \ 147 do { \ 148 const curandStatus_t _p_curand_stat__ = __VA_ARGS__; \ 149 if (PetscUnlikely(_p_curand_stat__ != CURAND_STATUS_SUCCESS)) { \ 150 if (((_p_curand_stat__ == CURAND_STATUS_INITIALIZATION_FAILED) || (_p_curand_stat__ == CURAND_STATUS_ALLOCATION_FAILED)) && PetscDeviceInitialized(PETSC_DEVICE_CUDA)) { \ 151 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU_RESOURCE, \ 152 "cuRAND error %d. " \ 153 "Reports not initialized or alloc failed; " \ 154 "this indicates the GPU has run out resources", \ 155 (PetscErrorCode)_p_curand_stat__); \ 156 } else { \ 157 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU, "cuRand error %d", (PetscErrorCode)_p_curand_stat__); \ 158 } \ 159 } \ 160 } while (0) 161 #define CHKERRCURAND(...) PetscCallCURAND(__VA_ARGS__) 162 163 PETSC_EXTERN cudaStream_t PetscDefaultCudaStream; // The default stream used by PETSc 164 PETSC_EXTERN PetscErrorCode PetscCUBLASGetHandle(cublasHandle_t *); 165 PETSC_EXTERN PetscErrorCode PetscCUSOLVERDnGetHandle(cusolverDnHandle_t *); 166 PETSC_EXTERN PetscErrorCode PetscGetCurrentCUDAStream(cudaStream_t *); 167 168 #endif // PETSC_HAVE_CUDA 169 170 // these can also be defined in petscdevice_hip.h so we undef and define them *only* if the 171 // current compiler is NVCC. In this case if petscdevice_hip.h is included first, the macros 172 // would already be defined, but they would be empty since we cannot be using HCC at the same 173 // time. 174 #if PetscDefined(USING_NVCC) 175 #undef PETSC_HOST_DECL 176 #undef PETSC_DEVICE_DECL 177 #undef PETSC_KERNEL_DECL 178 #undef PETSC_SHAREDMEM_DECL 179 #undef PETSC_FORCEINLINE 180 #undef PETSC_CONSTMEM_DECL 181 182 #define PETSC_HOST_DECL __host__ 183 #define PETSC_DEVICE_DECL __device__ 184 #define PETSC_KERNEL_DECL __global__ 185 #define PETSC_SHAREDMEM_DECL __shared__ 186 #define PETSC_FORCEINLINE __forceinline__ 187 #define PETSC_CONSTMEM_DECL __constant__ 188 #endif 189 190 #ifndef PETSC_HOST_DECL // use HOST_DECL as canary 191 #define PETSC_HOST_DECL 192 #define PETSC_DEVICE_DECL 193 #define PETSC_KERNEL_DECL 194 #define PETSC_SHAREDMEM_DECL 195 #define PETSC_FORCEINLINE inline 196 #define PETSC_CONSTMEM_DECL 197 #endif 198 199 #ifndef PETSC_DEVICE_DEFINED_DECLS_PRIVATE 200 #define PETSC_DEVICE_DEFINED_DECLS_PRIVATE 201 #define PETSC_HOSTDEVICE_DECL PETSC_HOST_DECL PETSC_DEVICE_DECL 202 #define PETSC_DEVICE_INLINE_DECL PETSC_DEVICE_DECL PETSC_FORCEINLINE 203 #define PETSC_HOSTDEVICE_INLINE_DECL PETSC_HOSTDEVICE_DECL PETSC_FORCEINLINE 204 #endif 205